Photorealistic talking faces from audio
Provided is a framework for generating photorealistic 3D talking faces conditioned only on audio input. In addition, the present disclosure provides associated methods to insert generated faces into existing videos or virtual environments. We decompose faces from video into a normalized space that decouples 3D geometry, head pose, and texture. This allows separating the prediction problem into regressions over the 3D face shape and the corresponding 2D texture atlas. To stabilize temporal dynamics, we propose an auto-regressive approach that conditions the model on its previous visual state. We also capture face illumination in our model using audio-independent 3D texture normalization.
1. A computing system to generate a talking face from an audio signal, the computing system comprising:
one or more processors; and
one or more non-transitory computer-readable media that collectively store:
a machine-learned face geometry prediction model configured to predict a face geometry based on data descriptive of an audio signal that comprises speech;
a machine learned face texture prediction model configured to predict a face texture based on data descriptive of the audio signal that comprises the speech; and
instructions that, when executed by the one or more processors, cause the computing system to perform operations, the operations comprising:
obtaining the data descriptive of the audio signal that comprises speech;
using the machine-learned face geometry prediction model to predict the face geometry based at least in part on the data descriptive of the audio signal;
using the machine-learned face texture prediction model to predict the face texture based at least in part on the data descriptive of the audio signal; and
combining the face geometry with the face texture to generate a three-dimensional face mesh model.
2. The computing system of claim 1 , wherein the audio signal comprises a standalone audio signal that is independent of a visual representation of the speech.
3. The computing system of claim 1 , wherein the data descriptive of the audio signal comprises a spectrogram of the audio signal.
4. The computing system of claim 1 , wherein:
the face geometry predicted based at least in part on the data descriptive of the audio signal is predicted within a normalized three-dimensional space associated with a three-dimensional mesh; and
the face texture predicted based at least in part on the data descriptive of the audio signal is predicted within a normalized two-dimensional space associated with a two-dimensional texture atlas.
5. The computing system of claim 1 , wherein the operations further comprise inserting the three-dimensional face mesh model into a two-dimensional target video to generate a synthesized video.
6. The computing system of claim 5 , wherein inserting the three-dimensional face mesh model into the two-dimensional target video comprises:
obtaining the two-dimensional target video;
detecting a target face in the two-dimensional target video;
aligning the three-dimensional face mesh with the target face at a target position; and
rendering the three-dimensional face mesh within the two-dimensional target video at the target position to generate the synthesized video.
7. The computing system of claim 5 , wherein inserting the three-dimensional face mesh model into the two-dimensional target video comprises:
generating a fixed atlas from the two-dimensional target video; and
providing the fixed atlas to the machine-learned face texture prediction model as a proxy lighting map.
8. The computing system of claim 6 , wherein detecting the target face comprises
using a three-dimensional face detector to obtain a pose and a triangle mesh of the target face in the video; and
decomposing the target face into a three-dimensional normalized space associated with a three-dimensional mesh and a two-dimensional normalized space associated with a two-dimensional texture atlas;
wherein the face geometry predicted based at least in part on the data descriptive of the audio signal is predicted within the normalized three-dimensional space associated with the three-dimensional mesh; and
wherein the face texture predicted based at least in part on the data descriptive of the audio signal is predicted within the normalized two-dimensional space associated with the two-dimensional texture atlas.
9. The computing system of claim 1 , wherein the operations further comprise rendering the three-dimensional face mesh in a three-dimensional virtual environment.
10. The computing system of claim 1 , wherein the audio signal comprises recorded human audio utterances or synthesized text-to-speech audio generated from textual data.
11. The computing system of claim 1 , further comprising performing lighting normalization on the face texture predicted by the machine-learned face texture prediction model.
12. The computing system of claim 1 , wherein the machine-learned face geometry prediction model and the machine learned face texture prediction model comprise personalized models that are specific to a speaker of the speech included in the audio signal.
13. The computing system of claim 1 , wherein the machine-learned face texture prediction model comprises an auto-regressive model that, for each of a plurality of iterations, receives as input a previous iterative prediction of the machine-learned face texture prediction model.
14. The computing system of claim 1 , wherein the predicted face texture comprises a combination of a difference map predicted by the machine-learned face texture prediction model and a reference texture atlas.
15. A computer-implemented method to learn to generate a three-dimensional face mesh from a training video, the method comprising:
obtaining, by a computing system comprising one or more computing devices, the training video comprising visual data and audio data, wherein the visual data depicts a speaker and the audio data comprises speech uttered by the speaker;
applying, by the computing system, a three-dimensional face landmark detector to the visual data to obtain three-dimensional facial features associated with a face of the speaker;
projecting, by the computing system and based on the three-dimensional facial features, the training video onto a reference shape to obtain a training face texture;
predicting, by the computing system and using a machine-learned face geometry prediction model, a face geometry based at least in part on data descriptive of the audio data;
predicting, by the computing system and using a machine-learned face texture prediction model, a face texture based at least in part on the data descriptive of the audio data;
modifying, by the computing system, one or more values of one or more parameters of the machine-learned face geometry prediction model based at least in part on a first loss term that compares the face geometry predicted by machine-learned face geometry prediction model with the three-dimensional facial features generated by the three-dimensional face landmark detector; and
modifying, by the computing system, one or more values of one or more parameters of the machine-learned face texture prediction model based at least in part on a second loss term that compares the face texture predicted by the machine-learned face texture prediction model with the training face texture.
16. The computer-implemented method of claim 15 , further comprising:
generating a fixed atlas from the training video; and
inputting the fixed atlas into the machine-learned face texture prediction model to serve as a proxy lighting map.
17. The computer-implemented method of claim 16 , wherein generating the fixed atlas comprises:
projecting the training video onto the reference shape using fixed reference facial coordinates; and
masking pixels that correspond to eye and inner mouth regions.
18. The computer-implemented method of claim 15 , wherein the machine-learned face texture prediction model comprises an auto-regressive model that, for each of a plurality of iterations, receives as input a previous iterative prediction of the machine-learned face texture prediction model.
19. The computer-implemented method of claim 15 , wherein the predicted face texture comprises a combination of a difference map predicted by the machine-learned face texture prediction model and a reference texture atlas.
20. One or more non-transitory computer-readable media that store instructions that, when executed by a computing system comprising one or more computing devices cause the computing system to perform operations the operations comprising:
obtaining, by a computing system, a training video comprising visual data and audio data, wherein the visual data depicts a speaker and the audio data comprises speech uttered by the speaker;
applying, by the computing system, a three-dimensional face landmark detector to the visual data to obtain three-dimensional facial features associated with a face of the speaker;
projecting, by the computing system and based on the three-dimensional facial features, the training video onto a reference shape to obtain a training face texture;
predicting, by the computing system and using a machine-learned face texture prediction model, a face texture based at least in part on the data descriptive of the audio data;
evaluating, by the computing system, a loss term that compares the face texture predicted by the machine-learned face texture prediction model with the training face texture; and
modifying, by the computing system, one or more values of one or more parameters of the machine-learned face texture prediction model based at least in part on the loss term.